audit-core/workplans/AUDIT-WP-0009-evidence-role-conformance.md

---
id: AUDIT-WP-0009
type: workplan
title: "Evidence-role conformance under Security Layer Model v0.7"
domain: infotech
repo: audit-core
status: active
owner: claude
topic_slug: railiance
created: "2026-08-29"
updated: "2026-09-11"
depends_on:
  - AUDIT-WP-0007
state_hub_workstream_id: "46a96b03-bc08-53b5-9c93-4071adabf734"
---

# AUDIT-WP-0009 — Evidence-role conformance under Security Layer Model v0.7

## Goal

Make audit-core dischargeable as the estate's **Evidence engine** under
NetKingdom Security Layer Model v0.7 (accepted), and stop it claiming a property
its own contract does not permit.

Two things drive this workplan.

**First, a live overclaim.** `audit_core/postgres_backend.py` returns
`tamper_evidence=True` as a constant, while `docs/integrity.md` permits that
claim only when a live external chain-head attestation exists. The one
attestation on record is 2026-08-16 and nothing renews it. Audit Core is
currently telling every sender it has a property whose precondition is
unverified — the §9.6 defect it twice corrected in gate-house's doctrine, turned
inward.

**Second, an obligation audit-core argued for and cannot yet support.** v0.7
§9.6 makes a declared emission cadence a **MUST** for load-bearing sources, with
reconciliation or a heartbeat as the required form for low-volume classes.
audit-core raised both, moving cadence from SHOULD to MUST. A source cannot
declare a cadence to a system with nowhere to put it. Having won the obligation,
audit-core owes the surface.

Full assessment: `history/2026-08-29-v0.7-alignment-and-scope-assessment.md`.
Standard: `net-kingdom/canon/standards/security-layer-model_v0.7.md`, companion
`net-kingdom/SECURITY-COMPANION.md`.

## Non-goals

Fixed by the statute; not deferred, not ours:

- No authorization decision surface, and no approval-validity query (§6, §9.4).
- No operative approval state (§9.4).
- No WORM, object lock, or ITC-CAP `data.archive` — §16 ruled the
  stronger-custody gap closed, a different store with a different owner.
- Emission atomicity at the source, which is the emitter's obligation (§9.6).

## Tasks

```task
id: AUDIT-WP-0009-T01
status: done
priority: high
state_hub_task_id: "f545b0e4-8c99-5186-affd-ce9a41209ed7"
```
Make `tamper_evidence` conditional on live attestation state. Derive the flag
rather than hard-coding it: the backend reports `True` only when a chain-head
attestation exists, verifies against the live chain, and is within a declared
freshness window. Stale or absent attestation degrades the claim — it does not
leave it standing. Assert the degradation with a test, and state the freshness
window in `docs/integrity.md` alongside the two existing preconditions, which
today are documented but unenforced.

Done 2026-09-06. `evaluate_tamper_evidence` in `audit_core/integrity.py` derives
the claim from a `ChainReport` and the mounted attestation; `PostgresAuditBackend`
reads it per `retention_policy` call through `tamper_evidence_state()`, briefly
cached because `/readyz` probes it. Seven degradation reasons are distinguished
and asserted — `no_attestation`, `attestation_stale`, `attestation_mismatch`,
`attestation_undated`, `chain_break`, `chain_unreadable`, against `attested`.
Unreadable is treated as absent deliberately: a malformed file must not hold up
a claim a missing file would drop. The freshness window is **168h** against an
intended daily cadence, declared in `docs/integrity.md` with the reasoning that
the window is seven cadences so missed runs degrade rather than flap. Twelve
tests in `tests/test_integrity.py`; `docs/audit-backend-contract.md` now states
the field is derived rather than declared.

Consequence to state plainly: production `/readyz` will report
`tamper_evidence: false` until T02 mounts a renewed attestation. The 2026-08-16
attestation is 21 days old against a 7-day window. That is the overclaim being
removed, not a regression — the claim was false before and now says so.

```task
id: AUDIT-WP-0009-T02
status: done
priority: high
state_hub_task_id: "6de9945f-4dd1-57dd-898a-f59c35b1df6c"
```
Done 2026-09-10. `deploy/attest-cronjob.yaml`: daily at 03:17 UTC against the
168h window, its own ServiceAccount, and a Role reaching exactly one named
ConfigMap — `get`/`update`/`patch`, no `create`, no `list`.

Three decisions worth recording, all of them refusals.

**The producer is not the receiver.** A receiver that could rewrite its own
attestation could forge it, so the job is a separate workload and the receiver
has no API-server egress. `audit-core-egress` is now scoped to
`component: receiver` and a new `audit-core-attest-egress` carries the 6443
rule; `tests/test_layer_conformance.py` asserts the receiver never gains it.

**It refuses to publish over a broken chain.** A fresh head written over a
break replaces an honest `chain_break` with a fresh-looking attestation. Stale
degrades the claim visibly; false does not.

**Mounted as a directory, not `subPath`.** Found while writing the manifest: a
`subPath` ConfigMap mount is resolved once at pod start and never updates, so
the daily attestation would land in the ConfigMap and never reach the running
receiver — `tamper_evidence` would age out to `false` while the job reported
success every night, silent in both directions.

`audit_core/attest_publish.py` does the publish in stdlib rather than shelling
out, because the image carries no `kubectl` and adding one to an audit
receiver's image to write a single file is the worse trade.

The offsite copy stays an operator step: audit-core holds no Nextcloud
credential and should not acquire one to publish a hash. Stated plainly in
`docs/integrity.md` — until that copy exists, the delivered control is "defends
against a database owner, not a cluster owner", and no stronger claim may be
made from it. Applying the manifests remains an operator action.
Schedule chain-head attestation so the precondition T01 enforces is normally
met. `attest-chain` exists and is operator-run; `deploy/` has no job. Add one,
write the attestation to the logical-offsite path already used by
`RESOURCE-WP-0002-T06` (`rapp-postgres` / Nextcloud + age) and **not** the
Barman prefix, per `docs/integrity.md` — that copy is restored with the table
and proves nothing. Record the cadence in `docs/integrity.md`.

```task
id: AUDIT-WP-0009-T03
status: done
priority: high
state_hub_task_id: "acae6085-ada7-51b2-8dc0-4abac7f7e7b3"
```
Represent the §9.6 evidence kind per source. Add `evidence_kind`
(`load-bearing` | `attributive`) to `SenderIdentity` and the sender registration
schema, defaulting to `attributive` — a source that has not declared is not
silently treated as load-bearing. Where a source declares `attributive` and has
deliberately traded away atomicity, carry the declared trade with it, because
§9.6 requires the trade be declared where the trail is documented. Prerequisite
for T04–T06.

Done 2026-09-06. `evidence_kind` and `completeness_trade` on `SenderIdentity`,
the `AUDIT_CORE_SENDERS` schema, and the non-secret scope overlay. Two
asymmetries are deliberate and asserted:

- **Default is `attributive`.** The other default would have audit-core imply a
  completeness obligation no source ever accepted.
- **The overlay may raise the kind, never lower it.** Same principle that stops
  an ExternalSecret refresh shrinking `user-engine`'s tenants: a ConfigMap
  refresh must not be able to drop a source's atomicity and detection
  obligations without anyone deciding to. Downgrading raises.

A load-bearing source may not carry a `completeness_trade` — §9.6 requires
atomicity of it, so recording a trade would store a contradiction as policy.
`evidence_declaration()` states `completeness_claimed: False` for **both**
kinds, because neither licenses the claim that the archive proves an event
occurred. Twelve tests in `tests/test_senders.py`.

```task
id: AUDIT-WP-0009-T04
status: done
priority: high
state_hub_task_id: "f36470af-4019-54b2-96d7-e049d867c7db"
```
Done 2026-09-10. `audit_core/stream_findings.py`, surfaced at
`GET /v1/stream-findings`, documented in `docs/stream-completeness.md`.

A heartbeat is an **ordinary event** — same envelope, same append-only custody,
same chain, no special table. A heartbeat stored outside the chain would be the
one record in this store that could be back-dated.

Declared **per class**, not per source, on the sender registration
(`heartbeat_classes`). `informed-decision` raised the shape first and it
generalises: a per-source heartbeat from a mixed-volume emitter is satisfied by
its chattiest class and says nothing about the quiet, security-relevant one,
which is the only reason heartbeats exist.

`no_heartbeat_since_registration` is a distinct finding kind rather than a skip
— a source that declared a heartbeat and never sent one is the case most likely
to be a broken integration, and it is exactly what a "compare against last
seen" implementation drops silently. A wildcard source is refused rather than
guessed at: there is no determinate set of streams to expect a heartbeat from.

Explicitly not T05's territory. Cadence describes expected *rate* and belongs
to the §17 Taxonomy schema; this says how often a source promises to say
*nothing to report* for a class that may legitimately be silent.
Heartbeat ingestion and missing-heartbeat findings. The required form for
low-volume load-bearing classes, and the only control covering adversarial
omission for exactly the rare negative events — revocation, denial, containment
— where suppression is most valuable and least visible. Accept a signed
*nothing to report* assertion per source and class; raise a finding when an
expected heartbeat does not arrive. The point is a positive claim that can
itself go missing, which rate monitoring can never produce.

```task
id: AUDIT-WP-0009-T05
status: wait
priority: medium
state_hub_task_id: "382575ca-1801-5a32-a93d-90a8b9c8adbf"
```
Accept and evaluate a declared emission cadence per source, and raise a finding
when the stream falls below it. **Waiting on** the §17 emission-cadence
declaration schema, which `kings-guard` is drafting and Taxonomy will own; do
not invent a competing shape. Cadence covers high-volume streams; T04 covers the
low-volume load-bearing case, and the two are complementary rather than
alternatives.

```task
id: AUDIT-WP-0009-T06
status: done
priority: medium
state_hub_task_id: "0a8d6eed-75dd-5aaa-bf86-60be9dadca03"
```
Done 2026-09-10. `GET /v1/reconciliation?source=&tenant=&since=&until=`,
per-class counts, `event_counts` on both the SQLite and Postgres backends.
Counts only — never payloads.

**The read-model tension, resolved rather than ignored.** The task said respect
`may_read`. But every registered sender holds `may_read: false`, so read
literally that makes the §9.6 reconciliation obligation undischargeable by
every source actually registered — a strange end for a rule audit-core argued
for. The resolution: a source asking how many of *its own* events audit-core
holds is not reading the archive, because it learns nothing it did not itself
emit. So the surface is scoped to the caller's own permitted sources and
tenants and carries no payloads; another source's counts, or an unscoped
aggregate, stay behind `may_read` and full tenant scope.

Two refusals: another source's counts return **403, not an empty count**, since
a zero would read as "we hold none of yours" — a false answer to a question
about completeness. And there is **no default window**: a count whose bounds the
caller did not choose is not comparable to anything the caller computed, and
would be quoted as though it were.
Reconciliation surface: per-source, per-class event counts over a bounded
window, readable by the source itself, so an emitter can compare audit-core's
count against its own state transitions and treat divergence as a finding.
`GH-WP-0002-T04` is the reference instance on the `approval-engine` side; this
is its counterpart here. Respect the existing read model — `may_read`, tenant
scoping, and no payloads in a counts response.

```task
id: AUDIT-WP-0009-T07
status: done
priority: medium
state_hub_task_id: "f572dfeb-0d1b-58d6-be83-405125189028"
```
Done 2026-09-10. `GET /v1/stream-findings`, following the existing
findings-surface conventions rather than adding a new shape: `may_read` plus
full tenant scope, because findings span every registered sender and carry no
tenant key to filter on — the same rule as dead letters, secret findings and
integrity.

INTENT principle 10 had the principle and no mechanism. It has one now.

Every finding and every reconciliation response carries a `means` field stating
the bound — a missing heartbeat is not proof of suppression, and agreement on
counts proves neither completeness nor that any event occurred. On the response
rather than in a document nobody opens beside it, because these are the numbers
most likely to be quoted out of context in someone else's conformance argument.
Give stream-completeness findings a home. `/v1/dead-letters` and
`/v1/secret-findings` exist; a cadence miss (T05) and a missing heartbeat (T04)
have nowhere to surface. INTENT principle 10 already says a degraded audit
stream is itself an audit and operations event — the principle is in place and
the mechanism is not. Follow the existing findings-surface conventions rather
than adding a new shape.

```task
id: AUDIT-WP-0009-T08
status: done
priority: medium
state_hub_task_id: "f7b513ee-af48-5c5a-a34e-3e9922d29b76"
```
Assert the §9.4 approval-validity prohibition with a negative test. It is
currently honoured by absence, which is not the estate's idiom: §6.4 obligation
3 requires a PEP's published stance map equal shipped behaviour *asserted by
test*. A prohibition worth stating in `INTENT.md`, `SCOPE.md`, and `layer.yaml`
is worth asserting in `tests/`.

```task
id: AUDIT-WP-0009-T09
status: progress
priority: high
state_hub_task_id: "fd4a4ac3-e525-57c1-9179-e0fcd0226913"
```
Register `approval-engine` as a distinct source under §9.4 and `AUDIT-IN-0001`:
sender registration, the four event classes (issuance, use, supersession,
revocation), tenancy mapping, retention profile, `secret_policy`, and
`evidence_kind: load-bearing` from T03. Onboarded under INTENT principle 4 —
declared ownership, retention, access, export and evidence policy, not merely
events arriving.

**Re-priced low → high, 2026-09-06.** `glas-harness` reports (message
`85c68e66`) that `approval-engine` production startup now *requires* audit
delivery, putting this on the deployment critical path. The old note — "not
blocking: `approval-engine` is not yet emitting" — read the dependency
backwards: audit-core is not waiting for a producer, the producer is waiting
for a sender identity. Corrected.

What this does not change: T09 still depends on **T03**, because §9.4 makes
approval evidence load-bearing and there is nowhere to record that until
`evidence_kind` exists. Registering the sender ahead of T03 would admit a
load-bearing source into a system that would silently treat it as attributive
— the precise defect §9.6 names. T03 is therefore promoted with it. Cadence
and reconciliation (T04/T06) remain open gates on the *detection* half and are
not prerequisites for admission; they bound what may be claimed of the stream
once it flows, and `approval-engine` should be told so rather than discovering
it later.

Token provisioning goes by the approved custody path
(`approval-engine/approval-engine-audit`, key `audit-token`) via `warden
route`; no sender token passes through Git, State Hub, or a workplan.

**In progress 2026-09-06.** T03 landed, so the registration inputs are now
expressible. Prepared, non-secret, and recorded in
`docs/approval-engine-source-registration.md`:

- Non-secret scope in `deploy/senders-scope.json` and `senders-scope.yaml`,
  `evidence_kind: load-bearing`, `may_read: false`, exact source match. Landed
  ahead of the credential deliberately: the overlay only applies to a sender
  the Secret already carries, so the entry admits nothing until the token
  exists. A test asserts that inertness rather than trusting the reading.
- `audit-core-approval-engine-ingress` in `deploy/networkpolicies.yaml`:
  namespace **and** pod label in one `from` peer, so both are ANDed. Narrower
  than `user-engine`'s namespace-only rule on purpose — a new sender should not
  inherit an older rule's breadth. `user-engine`'s policy is unchanged, and a
  test asserts it stays that way.
- The four §9.4 classes, retention (no expiry; 30-day `measured` recoverable
  window), custody class, and the claim bound.

**Two inputs are owed by `approval-engine`, not by audit-core.** `tenants` is
proposed as `["*"]` and needs confirmation with reasoning or a list —
`senders.py` says a missing tenant restriction must be justified per sender,
and audit-core cannot justify it on another repo's behalf. `secret_policy` is
recommended `redact` (losing a revocation record is worse than storing it with
one field redacted) but should be chosen explicitly rather than inherited.
Applying the manifests is an operator action audit-core does not take
unprompted.

**T02 is not a gate on this.** Asked directly by `glas-harness`: attestation
freshness governs what audit-core may *claim*, not whether it accepts and
durably stores. The append-only trigger and hash chain operate regardless. And
it is not even the relevant control — §9.6 is explicit that the acute risk for
approvals is omission, which tamper evidence does not address at all. T04 does.
A consumer waiting on T02 before trusting approval evidence would be waiting on
the wrong thing.

**2026-09-08 source return:** approval-engine now records the exact existing
store scope `["tenant:platform"]` and explicitly selects `redact` in
`docs/audit-source-registration.md`. This supersedes the two input waits above.
The JSON and ConfigMap overlays are narrowed together. The overlay preserves
secret_policy from the protected registry, so provisioning must set it there.
T09 remains progress for the platform custody record, credential, admitted
manifest application and positive/negative live ingestion evidence. No secret
was created and no production manifest was applied by this source change.

```task
id: AUDIT-WP-0009-T10
status: done
priority: low
state_hub_task_id: "ac3464fd-3df9-51a1-b1f4-3bf3c60e4265"
```
Done 2026-09-10. `layer.yaml` now lists four infrastructure contacts —
`platform-pg`, `state-hub`, `kube-apiserver`, and the container registry — each
with its role and whether another layer reads it. `tooling_contacts` stays `[]`,
which is true under §5 as written; the companion's totality request is met by
the uncatalogued list rather than by inventing a Tooling row.

`tests/test_layer_conformance.py` makes the totality claim mechanical rather
than a promise to remember: the egress destinations are derived from
`deploy/networkpolicies.yaml`, the registry from the pinned image digests, and
a new contact appearing in the manifests without a row here fails the test. It
also pins the declaration's own content — engine/evidence, no decision surface,
`approval_validity_query: forbidden`, and an evidence bound that never claims
occurrence.
Make the §5 conformance check total. `layer.yaml` declares
`tooling_contacts: []`, true under §5 as written — audit-core is an Engine and
holds no `key-cape` or OpenBao client. Companion §4 asks that uncatalogued
infrastructure be listed anyway, and that carve-out sunsets within two review
intervals for a store another layer reads, which `platform-pg` is. List it and
the State Hub now. Add a conformance script and test in the shape ops-warden
offered estate-wide.

```task
id: AUDIT-WP-0009-T11
status: progress
priority: high
state_hub_task_id: "96e0cbd5-e3ce-5807-8f64-41244671327b"
```
Register `informed-decision` as a distinct load-bearing source under
`AUDIT-IN-0003`, and answer the doctrine question `GH-DEC-2026-014` put to
audit-core: whether the record shape can carry a source-held-content
declaration with a retrieval expectation.

`informed-decision` is the browser-facing approver surface. `GH-DEC-2026-012`
limit 3 makes its evidence copy the one that must reach audit-core
independently of the emitter, because there the actor being audited and the
evidence source are the same component. `GH-DEC-2026-014` then granted
commitment-only payloads for Stage 1 under a condition Gate House added: the
record must assert that committed content **exists** and where **custody**
sits, so non-production is a finding attributable to the custodian rather than
an unremarkable blank.

**Done 2026-09-10, prepared and inert:**
`docs/informed-decision-source-registration.md`, scope entries in
`deploy/senders-scope.{json,yaml}`, and
`audit-core-informed-decision-ingress` in `deploy/networkpolicies.yaml`
(namespace AND pod label in one peer, following approval-engine rather than
`user-engine`'s older breadth). Every proposed field accepted as-is: exact
source, `["tenant:platform"]`, write true, read false, `load-bearing`,
`secret_policy: redact`.

**The answer to Gate House's condition, in two halves that must travel
together.** The fabric *can* carry the declaration — `data` is stored verbatim
into `details.data` and chained, so `content_exists` and `custody` need no
schema change and become as tamper-evident as the commitment they accompany. It
*cannot* detect non-production: audit-core performs no retrieval and its egress
policy permits Postgres and DNS only, so it never learns whether a custodian
would produce. Detection happens at retrieval, by the reviewer, and the stored
declaration is what turns a blank into an attributable failure. Residual stated
rather than left to be found: a custodian that never held the content can emit a
false `content_exists`; audit-core validates the declaration's shape, never its
truth. Not a requirement nobody can meet, and not the stronger thing it might be
mistaken for.

**Cadence answered:** reconciliation-per-class plus heartbeat for the low-volume
classes is the right shape for a mixed-volume source, with both scoped **per
class** rather than per source — a per-source heartbeat is satisfied by the
high-volume presentation stream and says nothing about a quiet month of
dispositions. Bound stated: where the emitter itself is compromised, it
suppresses the event and its own count together, so neither control covers the
residual `GH-DEC-2026-012` limit 3 exists for. Depends on T04/T06 for the
detection half; declaring the cadence now is correct, describing it as operating
is not.

Remaining: token custody, the protected registry entry, operator manifest
application, and live ingestion evidence. No secret was created and no
production manifest applied.


### Factory sender deployment precondition — 2026-09-11

Native metadata still reports 1/1 Ready on image `c2fe39a0185b...`. The exact
image and a synthetic SenderRegistry probe in the native pod confirm that it
has no `evidence_kind` support. Current source supports the contract. The
source-complete T03 implementation has not reached this deployment; readiness
alone cannot admit a load-bearing sender. T09/T11 retain the compatible image
release, current scope/peer manifests and actual ingestion/retrieval acceptance.

Platform RPF-WP-0035-T08 now prepares the separate sender custody records
CCR-2026-0021 (approval-engine) and CCR-2026-0022 (informed-decision), with
exact-path ESO projections and CAS-preserving first provision/resume. Its
receiver check refuses the current image before credential access. Seventeen
new tests, including eleven local OpenBao cases, and 53 existing credential-change
tests pass. The user approved both named owner review sets on 2026-09-11.
Compatible receiver release and target namespace readiness remain prerequisites.
See railiance-platform/docs/credential-lane-designs/factory-audit-senders-review.md
and its dated preparation evidence. Existing sender/source scopes are preserved.

## Acceptance

- No custody claim is returned unconditionally where `docs/integrity.md`
  conditions it, and a stale attestation visibly degrades the claim.
- A source can declare its evidence kind, and a load-bearing source can
  discharge the §9.6 cadence MUST against a real audit-core surface.
- A suppressed low-volume load-bearing class produces a finding rather than
  silence.
- The approval-validity prohibition is asserted by test, not by absence.
- `layer.yaml` lists every infrastructure contact, catalogued or not, and a test
  checks it.

## Notes

T01 and T02 are separable and T01 is the one that matters: scheduling
attestation without making the claim conditional would leave the overclaim in
place whenever the job fails, which is precisely the case the claim is supposed
to cover.

T05's dependency is real and should not be worked around. audit-core argued the
cadence obligation up to a MUST; inventing a local cadence shape ahead of the
Taxonomy schema would fragment the very interface it asked for.

### Compatible receiver release rehearsal — 2026-09-11

229 receiver/contract tests pass with real disposable PostgreSQL. The full suite
exposed five historical SQLite-import failures, now handed off as AUDIT-IN-0004;
that importer is not used by the native receiver. No migration SQL changed.
The new image uses the factory Alpine baseline and hashed runtime dependencies.
Candidate scan went from 81 High/3 Critical to zero reported vulnerabilities,
with the Alpine lifecycle-list warning retained. SIGTERM previously killed
PID 1 with exit 137; explicit handling now exits 0, preserves one in-flight
write and returns duplicate on restart/retry. Both synthetic senders pass exact
scope, redaction, independent readback and seven read-route denials.

[Rehearsal evidence](../docs/evidence/2026-09-11-factory-receiver-release.json).
Publish and native rollout are the next steps; T09/T11 remain progress until
real sender custody, delivery and owner acceptance are evidenced.